Autoencoder Reconstruction Residuals

Autoencoders learn the manifold of normal operation and flag anything they cannot faithfully reconstruct, catching novel precursors without labeled failures.

Unsupervised by necessity

Pre-FOAK there are few real failure examples, so Kronos cannot rely on supervised classifiers alone. Autoencoders sidestep this: they are trained only on normal operation to compress and reconstruct the feature vector. When the input lies on the learned manifold of normal behavior, reconstruction is accurate; when it does not, the reconstruction error rises. That error is the anomaly score.

# reconstruction-residual anomaly score
z     = encoder(x)              # compress to latent
x_hat = decoder(z)             # reconstruct
res   = mahalanobis(x - x_hat, Sigma_normal)   # whitened residual
score = res                    # high => off-manifold => anomalous

Kronos whitens the residual by the covariance of normal reconstruction error so that noisy-but-normal channels do not dominate the score. Variational and denoising variants make the latent well-behaved and force the model to learn robust structure rather than memorize noise, which matters given the high-neutron-flux electrical environment.

Localizing the anomaly

Because the residual is per-channel, the autoencoder localizes: it says not only that something is off but which signals are least reconstructable, e.g. a cluster of REBCO strain and voltage channels on one magnet section, or the end-cell diagnostics of the burner plug. That localization feeds directly into the ensemble's aggregation and into operator displays.

Guarding against normalization of failure

An autoencoder trained on data that already contains slow degradation will learn to call degradation normal. Kronos guards this by training on curated, validated-healthy periods and by monitoring the autoencoder's own baseline residual over time; a rising healthy-state residual signals model drift and triggers retraining at L0 rather than silent desensitization.